My bet is on some Fourier transform fuckery

I had to do a bit of research to remind myself. We did this in university as an exercise, so it was a while ago.

Basically with N signals (speakers) and M mixes (microphones). When M > N then you can create an MxN matrix and solve it numerically, and we were able to do that pretty easily. We also did the case where M=N as a demo and we used ICA for that with either projection pursuit or SVD, I can't remember.

You can read more about this on Wikipedia:

• Blind Source Separation
• Independent Component Analysis
• The exercise we did was a lot like this tutorial https://arxiv.org/pdf/1404.2986

With modern models this would be even easier with fewer microphones, and there's been a lot of time and money between when these methods were devised and now.